Domain Generalization-Based Damage Detection in Real Composite Structures Using Only Multiphysics Simulation Data

Structural health monitoring using ultrasonic guided waves has been widely applied to carbon fiber reinforced plastics (CFRP) for data-driven damage detection. However, acquiring monitoring signals experimentally is costly and destructive. To address this, we integrate multiphysics simulation with domain generalization to propose a method that enables effective damage detection on real signals while training exclusively on virtual signals. Finite element models are developed to simulate guided wave propagation in composite structures, incorporating a specially designed damage injection strategy to generate high-fidelity virtual signals. We propose a domain invariant and contrastive network (DICN) trained within a knowledge distillation framework, which explores both internal and mutual invariance of damage-related features. This enables DICN to learn discriminative representations from virtual signals and generalize to real signals. Experiments on CFRP laminates with different ply orientations show that DICN achieves over 80% accuracy in damage detection without real signal training, outperforming existing methods.